introduce trilinear upsampling

.buildkite/pipeline.yml

@@ -8,3 +8,13 @@ steps:
       queue: "juliagpu"
       cuda: "*"
     timeout_in_minutes: 60
+
+  - label: "GPU integration with julia v1.7-nightly"
+    plugins:
+      - JuliaCI/julia#v1:
+          version: "1.7-nightly"
+      - JuliaCI/julia-test#v1: ~
+    agents:
+      queue: "juliagpu"
+      cuda: "*"
+    timeout_in_minutes: 60

.gitignore

@@ -0,0 +1 @@
+Manifest.toml

Project.toml

@@ -1,6 +1,6 @@
 name = "NNlibCUDA"
 uuid = "a00861dc-f156-4864-bf3c-e6376f28a68d"
-version = "0.1.0"
+version = "0.1.7"
 
 [deps]
 CUDA = "052768ef-5323-5732-b1bb-66c8b64840ba"
@@ -10,8 +10,8 @@ Random = "9a3f8284-a2c9-5f02-9a11-845980a1fd5c"
 Statistics = "10745b16-79ce-11e8-11f9-7d13ad32a3b2"
 
 [compat]
-CUDA = "3"
-NNlib = "0.7"
+CUDA = "3.3.1"
+NNlib = "0.7.25"
 julia = "1.6"
 
 [extras]

README.md

@@ -1,5 +1,5 @@
-# NNlibCUDA
+# NNlibCUDA.jl
 
-CUDA compatibility for NNlib.jl.
+This is a glue package which extends functions from [NNlib.jl](https://github.com/FluxML/NNlib.jl) to work with [CUDA.jl](https://github.com/JuliaGPU/CUDA.jl). It should be loaded automatically when using [Flux.jl](https://github.com/FluxML/Flux.jl), but not when using NNlib.jl by itself.
 
 Julia gpu kernels are in `src/`, while wrappers around `cudnn` are in `src/cudnn/`.

src/NNlibCUDA.jl

@@ -4,9 +4,14 @@ using NNlib
 using CUDA
 using Random, Statistics
 
+const IntOrIntTuple = Union{Integer, NTuple{N,<:Integer} where N}
+
 include("upsample.jl")
 include("activations.jl")
 include("batchedmul.jl")
+include("scatter.jl")
+include("gather.jl")
+include("utils.jl")
 include("cudnn/cudnn.jl")
 include("cudnn/conv.jl")
 include("cudnn/pooling.jl")